Fault Detection via Sparse Representation for Semiconductor Manufacturing Processes

Abstract

Metal etching process is considered to be a bottleneck step in semiconductor manufacturing, and many parameters of etchers can be monitored with sensor systems. So, an effective fault detection algorithm for the etching process is particularly important. Recently, some classification methods that only need positive samples such as principal component analysis-based methods and k-nearest neighbor rule based methods have been used for fault detection and achieved some meaningful results. But the requirement that the distribution of the monitored parameters obeys certain structure is hard to satisfy due to product mix. Our method-fault detection via sparse representation is based on the assumption that there exists a similar pattern to the testing sample in the training set, while the fault is exotic to the training set. So, normal samples can be linearly represented well (a low representation error) using few training samples while the fault cannot reach the same error precision, which is independent of the distribution of the samples. Once the representation of the testing sample is determined using sparse representation model, the distance between the original sample and its representation using few high-weighted basis can be used to detect the fault. The illustrative and industry experiments have verified the efficiency of our method.